A Temperature Estimation Method for Simulated Annealing Using Neural Networks

Although simulated annealing is powerful to obtain good approximation solutions for multi-modal function optimization, it requires large amount of computation time. In order to speed up simulated annealing, it is worth while providing a good initial solution using high speed greedy heuristics, and then starting simulated annealing. One of the difficulties of this approach is the determination of the starting temperature of sim ulated annealing. The starting temperature should be such temperature that simulated annealing does not destroy some good structures of the initial solution produced by heuristic and ensures further optimization. We propose a solution to this problem, a new temperature estimation method for simulated annealing using neural networks. In the proposed method, the neural network learns some relationship between tempera ture and probability distribution of the difference of cost function between two neighboring solutions in the solution space. The neural network learns to estimate proper annealing temperature of solutions which are produced by greedy heuristics. We demonstrate the efficiency of the proposed method by applying it to VLSI cell placement problems. Computational experiments show that the proposed method can obtain the same quality placement results as those of simulated annealing within less computation time.